Académique Documents
Professionnel Documents
Culture Documents
--------------------------------------------------------
A Dissertation Presented to
University of Houston
--------------------------------------------------------
In Partial Fulfillment
Doctor of Philosophy
-------------------------------------------------------
By
Carlos Cobos
December 2014
MARKOV CHAIN STOCHASTIC SEISMIC INVERSION FOR
____________________________________
Carlos Cobos
APPROVED:
____________________________________
____________________________________
____________________________________
____________________________________
___________________________________
ii
ACKNOWLEDGMENTS
outstanding advisor, who has provided assistance, guidance and support. John is
much more than an academic advisor; he is truly a “maestro”. I would also like to
thank Aibing Li, Evgeny Chesnokov and Franklin Ruiz, for their readiness to
serve on my oral exam committee and for their thorough review and constructive
comments on my dissertation and for their tremendous help and support to finish
it. Special thanks go to Aibing Li for her help and guidance as graduate student
advisor in Geophysics.
I want to thank Sylvia Marshal for her patience and help in all these great
years. I am also very grateful to Anne Decour and Hua-wei Zhou; without their
I have enjoyed and learned from many discussions with Igor Escobar,
Tapan Mukerji and De-Hua Han; I cannot thank them enough for all the support
friends. Among them the most memorable are Carlos Moreno, Javier Villegas,
Tony De Lilla, Maria Villegas, Alexandra Marquez, Carlos Lopez, Karen Romero,
iii
Laszlo Benkovics, Ana Paredes, Freddy Obregon Roberto Varade and Noemi
Ordonez.
Carlos Eduardo, my siblings, Gracia, Carlos Jose, Ligia and Francisco Jesus,
and the rest of my family for their continued trust and support.
Alejandra, Carla and Carlos Eduardo. They have been the energy and motivation
iv
MARKOV CHAIN STOCHASTIC SEISMIC INVERSION FOR
---------------------------------------------------
An Abstract Dissertation
Presented to
University of Houston
--------------------------------------------------------
In Partial Fulfillment
Doctor of Philosophy
-------------------------------------------------------
By
Carlos Cobos
December 2014
v
ABSTRACT
generate elastic properties instead of the key rock properties needed to reliably
lithology, pore structure and fluid variations. The result of the probabilistic seismic
inversion is a seismic lithofacies catalog that can describe the elastic response of
the studied subsurface interval. The main advantage of this technique is that the
geologic scenarios and physical conditions, with the prime goal of minimizing
produces multiple equally probable rock property models that can explain the real
vi
and well-log data to demonstrate the feasibility of the estimation of critical rock
fraction. The real-data test results confirmed the capability of the proposed
vii
CONTENTS
viii
5.2. Synthetic Test 1: Cemented Sand ................................................. 52
5.3. Synthetic Test 2: High Porosity Sands .......................................... 66
5.4. Conclusions................................................................................... 77
6. 1D Real Data Application ................................................................ 79
6.1. Introduction............................................................................... 79
6.2. Seismic Data Conditioning ....................................................... 81
6.3. The Model Space Generation................................................... 83
6.4. The Seismic Stochastic Inversion ............................................. 85
6.5. Analysis .................................................................................... 90
6.6. Conclusions .............................................................................. 95
7. 3D Real Data Application ................................................................ 96
7.1. Introduction............................................................................... 96
7.2. The Kriging Estimator ............................................................... 96
7.3. 3D Real Data Application ......................................................... 99
7.4. Convergence Criteria ............................................................. 109
7.5. Algorithm Efficiency ................................................................ 111
6.7. Conclusions ............................................................................ 112
References ............................................................................................ 113
ix
LIST OF TABLES
x
LIST OF FIGURES
Figure 2.1: Probability density function of the discrete random variable described
in Table 2.1. .......................................................................................................... 9
Figure 2.2: Cumulative probability density function of the RV described in Table
2.2. ..................................................................................................................... 11
Figure 2.3: Schematic representation of the transition probability matrix. .......... 15
Figure 2.4: Markov Chain evolution after 500, 1,000, 5,000, 10,000, 25,000 and
50,000 steps. The equilibrium distribution is a Gaussian with zero mean and
variance of one. .................................................................................................. 18
Figure 2.5: Two-dimensional RW with 10,000 points.......................................... 19
Figure 3.1: Deterministic (left) versus stochastic models (right). The stochastic
model generates a set of possible values instead of a simple outcome. ............ 25
Figure 3.2: P-wave velocity histograms for two different lithofacies.................... 26
Figure 3.3: Three-dimensional projections of a ten-dimensional Gaussian PDF
sampled with 1,000 (upper left), 10,000 (upper right), 100,000 (lower left) and
1,000,000 (lower right). ....................................................................................... 29
Figure 3.4: 2D Projections of the multivariate probability distribution function. A)
Bivariate P-wave velocity versus S-wave velocity PDF; B) Bivariate P-wave
velocity versus bulk density PDF; C) Bivariate S-wave velocity versus bulk
density PDF; and D) Bivariate Porosity versus bulk density PDF. ...................... 31
Figure 3.5: 2D Projections of the modeled multivariate probability distribution
function. A) Bivariate P-wave velocity versus Vs PDF; B) Bivariate P-wave
velocity versus bulk density PDF; C) Bivariate S-wave velocity versus bulk
density PDF; and D) Bivariate Porosity versus bulk density PDF. ...................... 33
Figure 3.6: Elastic properties comparison between original saturation (left) and
gas replaced saturation (right). ........................................................................... 34
Figure 3.7: Example of theoretical rock physics models application to measured
P-wave velocity samples. The fluid replacement, cement modeling and porosity
perturbation effects are shown by red, green and black arrows. ........................ 35
Figure 3.8: Example of theoretical rock physics models application to measured
P-wave velocity samples. The probability distribution functions for the fluid
replacement, cement modeling and porosity perturbation effects are in red, green
and black, respectively. ...................................................................................... 35
Figure 3.9: Comparison of an original bivariate distribution (left) and the MCMCS
distribution (right). ............................................................................................... 37
xi
Figure 3.10: Mean square error plot for the MCMCS of the example shown in
Figure 3.9. .......................................................................................................... 37
Figure 4.1: The inverse and forward problems are complementary mathematical
problems. ............................................................................................................ 38
Figure 4.2: Simplified 1D stochastic inversion workflow. .................................... 41
Figure 4.3: Schematic representation of the possible scenario generation. ....... 43
Figure 4.4: Comparison among the depth, time and regularized time 1-D P-wave
velocity arrays. The black circles show how the regularization in the time domain
can affect the elastic properties of the modeled subsurface. .............................. 46
Figure 4.5: Schematic representation of the synthetic traces generation
algorithm. ............................................................................................................ 48
Figure 4.6: Misfit calculation example. In this synthetic example the cumulative
L2-norm shows the intervals with larger misfit. .................................................... 49
Figure 5.1: Model 1 synthetic well logs showing (from left to right) facies
indicator, shale volume, P-wave velocity, bulk density, S-wave velocity, porosity,
P-wave impedance and S-wave impedance. ...................................................... 52
Figure 5.2: Synthetic model 1 pore fluids velocity and density profiles. .............. 54
Figure 5.3: Final seismic lithofacies multivariate PDF projections for the Vp-Vs
(Upper), Vp-Rhob (Center), and PhiT-Vp (Lower) spaces. Green, blue, brown
and red contours denote shales, brine sands, oil sands and gas sands,
correspondingly. Some empirical models are plotted for comparison including
Castagna’s mudrock line, Castagna’s 1993, Han’s 1986, Gardner’s sand,
Gardner’s shale, Voigt bound, Reuss bound and Nu’s modified Voigt bound for
quartz mixture. .................................................................................................... 55
Figure 5.4: Model 1 angle dependent reflectivity profiles.................................... 56
Figure 5.5: Ricker wavelet with a central frequency of 25 Hz. ............................ 57
Figure 5.6: Synthetic angle gather generated from the synthetic model 1. ......... 57
Figure 5.7: 2D Projections of the multivariate probability distribution function for
sands (left) and shales (right) calculated from the synthetic model 1. ................ 58
Figure 5.8: Real seismic trace (left) and possible models traces (right)
comparison. The selected traces have L2-norms smaller than 0.9. .................... 59
Figure 5.9: Synthetic versus modeled waveform comparison for four possible
models with α<0.9. ............................................................................................. 60
Figure 5.10: Probable models with α<0.9. The black, red, yellow and green lines
represent the top of upper sand, base of upper sand, top of lower sand and base
of lower sand, respectively. ................................................................................ 61
Figure 5.11: Sandstone fraction for probable models with α< 0.9....................... 62
xii
Figure 5.12: Total porosity of the probable models with α<0.9. The black, red,
yellow and green lines represent the top of upper sand, base of upper sand, top
of lower sand and base of lower sand, respectively. .......................................... 63
Figure 5.13: Volume of shale probable models with α<0.9. The black, red, yellow
and green lines represent the top of upper sand, base of upper sand, top of lower
sand and base of lower sand, respectively. ........................................................ 63
Figure 5.14: Effective porosity probable models with α<0.9. The black, red,
yellow and green lines represent the top of upper sand, base of upper sand, top
of lower sand and base of lower sand, respectively. .......................................... 64
Figure 5.15: Elastic models with α<0.9. The black, red, yellow and green lines
represent the top of upper sand, base of upper sand, top of lower sand and base
of lower sand respectively. ................................................................................. 64
Figure 5.16: Normalized histograms comparison between the rock properties of
the synthetic model and the probable solutions, for sand’s shale volume (upper
left), sand’s total porosity (lower left), shale’s shale volume (upper right) and
shale’s total porosity (lower right). Red and blue bars correspond to synthetic
model and solutions, respectively. ...................................................................... 65
Figure 5.17: Probable models with α<0.045. The black, red, yellow and green
lines represent the top of upper sand, base of upper sand, top of lower sand and
base of lower sand, respectively. ........................................................................ 66
Figure 5.18: Model 2 synthetic well logs showing (from left to right) facie
indicator, shale volume, porosity, P-wave velocity, bulk density and S-wave
velocity. Black, blue and red lines correspond to in situ, brine and gas
saturations. ......................................................................................................... 68
Figure 5.19: Synthetic model 2 pore fluids velocity (left) and density profiles
(right). Blue and red lines correspond to brine and gas, respectively. ................ 68
Figure 5.20: Final seismic lithofacies multivariate PDF projections for the Vp-Vs
(Upper), Vp-Rhob (Center), and PhiT-Vp (Lower) spaces. Grey, blue, and red
contours denote shales, brine sands and gas sands, correspondingly. Some
empirical models are plotted for comparison, including Castagna’s Mudrock line,
Castagna’s 1993, Han’s 1986, Gardner’s sand, Gardner’s shale, Voigt bound,
Reuss bound and Nu’s modified Voigt bound for quartz mixture. ....................... 69
Figure 5.21: Model 2 angle-dependent reflectivity profiles. ................................ 70
Figure 5.22: Synthetic angle gather generated from synthetic model 2. ............. 71
Figure 5.23: 2D Projections of the multivariate probability distribution function for
gas sands (left), shales (center) and brine sands (right), calculated from synthetic
model 2. .............................................................................................................. 71
Figure 5.24: Real seismic trace (left) and possible models traces (right)
comparison. The selected traces have L2-norms smaller than 0.88. .................. 72
xiii
Figure 5.25: Synthetic versus modeled waveform comparison for four possible
models with α<0.88. ........................................................................................... 73
Figure 5.26: Probable models with α<0.88. The green, black and white represent
the shales, gas sands and brine sandstone seismic lithofacies, respectively. .... 74
Figure 5.27: Depth of top of the sand (left), gas-water contact (center) and base
of the sand (right) histograms. ............................................................................ 74
Figure 5.28: Total porosity of the probable models with α<0.88. ........................ 75
Figure 5.29: Volume of shale probable models with α<0.88. .............................. 76
Figure 5.30: Effective porosity probable models with α<0.88 calculated from the
total porosity and shale volume properties. ........................................................ 76
Figure 5.31: Elastic models with α<0.88. ............................................................ 77
Figure 6.1: Well 1 logs for in situ saturation. ....................................................... 80
Figure 6.2: Near angle stack; the top of the upper and lower sands are located at
72 ms and 104 ms, respectively. ........................................................................ 80
Figure 6.3: Angle gather at the Well-1 location (Upper). The amplitude extraction
at the top of the upper sand shows a strong AVO effect (Lower). ...................... 81
Figure 6.4: Well logs converted to time showing the well to seismic tie (right
track); the black line corresponds to the measured seismic trace and blue the
synthetic seismogram. ........................................................................................ 82
Figure 6.5: Angle-dependent extracted wavelets................................................ 83
Figure 6.6: Well 1 logs after fluid substitution. Black, red and blue lines
correspond to in situ, 90% gas and 100% brine-saturated sands. ...................... 84
Figure 6.7: 2D Projections of the multivariate probability distribution function
generated from Well 1 log data and Gassmann fluid substitution. ...................... 84
Figure 6.8: Comparison between real angle traces (upper and lower left) and the
320 solutions for 0° (upper right) and 32° (lower right) incident angles. ............. 86
Figure 6.9: Probable models with α≤0.88. Grey, black and white represent
shales, gas sands and low saturation gas sands, respectively. .......................... 87
Figure 6.10: Total porosity of the 320 probable models. ..................................... 88
Figure 6.11: Volume of shale of the 320 probable models. ................................ 88
Figure 6.12: Effective porosity of the 320 probable models. ............................... 89
Figure 6.13: Possible elastic models with α≤0.88. .............................................. 89
Figure 6.14: Probability of sand vertical plot. ...................................................... 91
Figure 6.15: Total porosity and effective porosity comparison. The black lines
correspond to the well log information and the blue lines to the inverted rock
properties............................................................................................................ 92
xiv
Figure 6.16: Most expected fluid seismic lithofacies (left) and probability of fluid
type (right). ......................................................................................................... 93
Figure 6.17: Comparison between the most expected elastic models (blue line)
against well logs (black lines). ............................................................................ 94
Figure 7.1: Simplified 3D stochastic inversion workflow. .................................. 100
Figure 7.2: Anisotropic semivariogram model with a 0.1 nugget, a sill of 0.9 and
ranges of 500 meters and 50 meters. ............................................................... 100
Figure 7.3: Variance model evolution for one of the stochastic simulation
outcomes. ......................................................................................................... 102
Figure 7.4: Three possible stochastic models. Shales are represented by facies
1, gas sands by facies 2, LSGS by facies 3 and brine sands by facies 4. ........ 102
Figure 7.5: Three possible stochastic total porosity models. ............................ 103
Figure 7.6: Three possible stochastic volume of shale models......................... 103
Figure 7.7: Three possible stochastic P-wave impedance models. .................. 104
Figure 7.8: Three possible stochastic bulk density models. .............................. 104
Figure 7.9: Three possible stochastic S-wave velocity models. ........................ 105
Figure 7.10: Most expected total porosity model generated from 400 stochastic
models. ............................................................................................................. 106
Figure 7.11: Most expected volume of shale model generated from 400
stochastic models. ............................................................................................ 107
Figure 7.12: Most expected volume of P-wave velocity generated from 400
stochastic models. ............................................................................................ 107
Figure 7.13: Most expected volume of density generated from 400 stochastic
models. ............................................................................................................. 108
Figure 7.14: Most expected volume of S-wave velocity generated from 400
stochastic models. ............................................................................................ 108
Figure 7.15: Models randomly selected from the real dataset inversion. .......... 109
Figure 7.16: Comparison between real angle trace (right) and 150,000 solutions
(left) randomly selected from the five million realizations performed for the real
example inversion............................................................................................. 110
Figure 7.17: Algorithm efficiency, the proposed algorithm increase linearly with
the number of tested models. ........................................................................... 111
xv
1. INTRODUCTION
1.1. Motivation
developed more than forty years ago for exploration plays where hydrocarbon
and amplitude versus offset (AVO) inversion. DHIs plays are usually related to
becoming obsolete, since the oil industry is moving to explore areas in which the
reservoirs with nearly undetectable pore fluid response. This means that the
as the porosity, lithology and rock fabric of compacted and cemented porous
rock.
1
The second-most important SRC challenge is to improve seismic vertical
solutions are usually limited to the acoustic (P-wave impedance) or elastic (P-
wave impedance, S-wave impedance, and bulk density) properties of the studied
subsurface intervals. However, these properties are not directly useful for
on the Markov Chain Monte Carlo Simulation (MCMCS) that can be implemented
can describe the elastic seismic response of the studied interval and its
associated uncertainties.
2
The main advantage of this petrophysical inversion technique is that the
algorithm can support the evaluation for hydrocarbon potential considering the
scenarios and physical conditions, with the prime goal of minimizing uncertainties
and risk. In field development areas, the stochastic inversion produces many
equally probable rock property models that can explain the real 3D seismic
( ) ( ) ( ( )) (1.1)
function in the model space, ( ) is the data probability density function and
( ) is the forward modeling operator that relates the model space to the data
3
space. The operator ( ) is usually non-linear and its role is to predict the
( ) (1.2)
that directly produces rock property volumes instead of elastic property volumes;
hence, the inversion results can be directly introduced into the geological model
sampling algorithms.
space for the proposed inversion scheme. We show how probability density
4
functions are generated from the integration of the available well and geological
data, and are supported by theoretical rock physics models. Second we describe
how the MCMCS is used to populate the model space for the proposed
and its three main steps are explained: the rock properties scenario generation,
the synthetic data generation and the comparison between the real and the
inversion scheme on two different sets of synthetic data. The synthetic data sets
were generated based on real well data, with the goal of testing the inversion in
geological setting where the sands have larger impedance than that of the
porosity environment where the fluid property variations have a very strong effect
scheme on a real set of data that include seismic and well log information. The
5
target interval is made up of four seismic lithofacies: shales, gas-filled sands, low
saturation gas sands (LSGS) and brine sands. The objective of this inversion is
to predict total porosity, volume of shale, effective porosity and fluid type
analyzed the same data set used in Chapter 6 that corresponds to tertiary marine
6
2. MARKOV CHAIN MONTE CARLO SIMULATION
2.1.Introduction
very large. Before introducing the MCMCS algorithm we define some basic
statistical concepts in Section 2.2. These basic concepts are random variables,
probability density functions, cumulative density functions and the most important
(Papoulis 1984), any function g(x) of real variables can be a RV if g(x) has the
following properties:
7
c) The probability of the events {X=±∞} must be zero.
can assume any value within a given range. For instance, if we randomly select
several individuals from a population and measure their height, the results can
take on any value within a given range, maybe from 2 to 7 feet. In contrast,
discrete variables can only assume certain values within a defined range. For
example, when we flip a coin and count the possible outcomes, no matter the
number of flips, the possible outcomes can only be heads or tails. Because the
discrete RV.
that describes the likelihood for a given RV to take on a given value. The
The following example will illustrate some concepts related to the random
probabilities associated with the possible outcomes are shown in Table 2.1.
8
Table 2.1: Probability associated to the possible outcomes of random process.
Outcome Probability
1 0.05
2 0.25
3 0.4
4 0.25
5 0.05
set of possible outcomes and their associated probabilities shown in Table 2.1
are the PDF of the RV X (Isaaks and Srivastava, 1989); Figure 2.1 shows the
0.4
0.35
0.3
0.25
Probability
0.2
0.15
0.1
0.05
0
1 2 3 4 5
Outcome
Figure 2.1: Probability density function of the discrete random variable described
in Table 2.1.
9
Suppose the random variable X may take N different values; the PDF of
( ) . (2.1)
By definition, the integral of the PDF (the sum of discrete RVs) is called
the probability that a RV X has a value less than or equal to x. The CDF of a
{ } ∑ (2.2)
Table 2.2 shows the CDF of the RV X shown in Table 2.1. The CDF of the
Outcome CDF
1 0.05
2 0.3
3 0.7
4 0.95
5 1
10
1
0.9
0.8
Cummulative Probability
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
1 2 3 4 5
Outcome
are useful to describe the statistics of the analyzed random variable. These
parameters include expected value and standard deviation of the RV. The
{ } ∑ (2.3)
{ }
11
expected value, more commonly known as dispersion. A small variance indicates
that the outcomes are very close to the expected value. In contrast, a large
variance indicates that the data is very spread out around the mean. Variance
{ } ∑ (∑ ) . (2.4)
the stochastic process are based only on its present state. A random variable is
b) The outcome at any stage depends only on the outcome of the previous
stage; and
( | ) ( | ) (2.5)
12
move from one stage to another. The probability that an RV in state xi will change
( ) ( | ). (2.6)
being in a particular state at the current step and the transition probability from
( ) ∑ ( | ) ( ). (2.7)
( ) ( ) , (2.8)
where S(n) is the probability vector at the nth step, S(0) is the probability vector at
A Markov Chain can be illustrated using a simple example. Suppose that the
price variation of a particular share in the stock market can be predicted using a
Markov process and the possible price states are: “decrease”, “same” or
“increase”. Therefore, the price of the share tomorrow only depends on today’s
13
share price. Assuming that today the price variation is “same”, the transition
( ).
( ) ( ).
( ) ( )( ) ( ).
( ) ( )( ) ( )
14
Figure 2.3: Schematic representation of the transition probability matrix.
The summary of the Markov Chain for the first six days is shown in Table
2.3. However, if the initial price variation is “decrease” instead of “same” (see
Table 2.4), the Markov Chain reaches the same probability values at the sixth
day. Hence, on the sixth day, the Markov Chain reaches its stationary
distribution, where the probability values are independent of the initial state.
15
Table 2.4: Markov chain for initial price variation “decrease”.
this dissertation we use the MCS to explore the space of possible solutions used
means that the quality of the MCMC distribution improves with the number of
steps. A Markov Chain is an ideal method for Monte Carlo simulations due to the
16
memoreless nature of Markov stochastic processes. This Monte Carlo method
properties;
and
reach the equilibrium with tolerable error. The better the Morkov Chain, the faster
the mixing time. Figure 2.4 shows an example of a Markov Chain with a
17
500 Samples 1000 Samples
40 60
50
30
40
Counts
20 30
20
10
10
0 0
-4 -2 0 2 4 -4 -2 0 2 4
250 500
200 400
Counts
150 300
100 200
50 100
0 0
-4 -2 0 2 4 -4 -2 0 2 4
3000
1000
Counts
2000
500
1000
0 0
-4 -2 0 2 4 -4 -2 0 2 4
Property Property
Figure 2.4: Markov Chain evolution after 500, 1,000, 5,000, 10,000, 25,000 and
50,000 steps. The equilibrium distribution is a Gaussian with zero mean and
variance of one.
using a stochastic method; coin tosses, Brownian motion and stock prices are
18
biased compared to systematic sampling algorithms. However, RWs can lead to
RW, the location can only jump to neighboring sites of the lattice. In a simple
location jumping to each one of its immediate neighbors are the same.
introduced in 1953 by a group of scientists (Metropolis et. al., 1953) from Los
19
Alamos National Laboratory, and extended to its general form by W. K Hasting in
2. Using the current value (xt) draw, generate the candidate value (xt+1) from
3. From the candidate (xt+1) and the current (xt) points, calculate the ratio of
( ) ( )
; (2.9)
( ) ( )
If α≥1 accept the candidate point (xt+1) as current value and return to step
2; otherwise, accept the candidate with probability α, else reject it and return to
step 2.
( ) ( )
( ). (2.10)
( ) ( )
20
As with any MCMC methods, the Metropolis–Hastings algorithm
MCMC sampler is the number of runs (steps) until the chain approaches
stationarity (the length of the burn-in period). Typically, the first 1,000 to 5,000
elements are thrown out, and then one of the various convergence tests is used
increase the required burn-in time, and an area of much current research is
whether an optimal starting point and proposal distribution can be found. One
suggestion for a starting value is to start the chain as close to the center of the
parameter space for long periods of time, as opposed to a well-mixed chain that
21
3. THE STATISTICAL ROCK PHYSICS MODEL
3.1.Introduction
and illustrates how well-log data and theoretical rock-physics models are used to
organic matter that are often amalgamated by cement minerals. The free space
situ conditions, the intergranular space is filled by liquid or gaseous fluids. The
solid portion of the rock is known as the rock’s matrix, or more commonly, “the
matrix”. The rock’s matrix geometry or how the crystalline grains are arranged in
of most sedimentary rocks and their elastic moduli; among these parameters, we
have volume fractions of minerals, nature of boding on the grain interface, type
properties (e.g., porosity, mineral and cement composition pore geometry, elastic
22
moduli, etc.). Rock-physics models are used to estimate rock properties that
have not been sampled. Most rock-physics models can be classified into three
theory (Biot, 1956) is one of the first to model the coupled mechanical behavior of
a porous rock filled with a linearly viscous fluid. At zero frequency, the Biot
experimental observations. Most empirical models are only valid for the specific
data sets and physical conditions where the data was measured. Empirical rock
physics models are widely used to predict the properties of rock for which
mudrock line (Castagna et. al., 1985), Gardner’s density relation (Gardner et. al,
experience-based techniques for problem solving. Since these models are non-
models are the Wyllie time-average equation (Wyllie et. al,. 1956) and the
23
3.3. Deterministic Versus Statistical Models
models, a given input will always produce the same output. In deterministic
24
Figure 3.1: Deterministic (left) versus stochastic models (right). The stochastic
model generates a set of possible values instead of a simple outcome.
Probabilistic rock physics models are physical models that relate the rock
of the rocks that constitute the sedimentary unit. Since these rocks have similar
statistics of the rock’s properties for any particular seismic lithofacies (Avseth,
2000).
25
We usually observe data scatter when we plot elastic properties against
rock properties for a given seismic lithofacies. Due to data scatter, descriptive
statistics are often used to summarize the elastic properties of a given seismic
properties of the statistical populations. Figure 3.2 shows the P-wave velocity
lithofacies A has a lower most expecting value and larger velocity dispersion than
the tendency of the data (Table 3.1), for example, when most of the data is
26
Table 3.1: Vp statistical summary table for seismic lithofacies A and B.
represent the complexity of the relations among rock properties, since the
definition of central estimators (like the mean or the median) and the dispersion
estimators (like the variance and standard deviation) generally do not have clear
(PDFs) are introduced to better describe the dependence among rock properties,
are generated from the integration of the available well and geological data and
are supported by theoretical rock physics models. In fact, multivariate PDFs are
usually hybrid rock physics models that can describe measured rock samples
A poorly mixed chain can arise because the target distribution is complex
(multi-modal) and our choice of starting values traps the Markov Chain to one of
27
the modes. An approach to solve this problem is to transform the complex target
then proceed with the MCMCS. This process is called normal scoring and is
To solve this issue, the number of samples has to be largely increased with the
1,000,000 draws. We can see that the distribution sampled with 1,000 draws
does not represent the shape of the target Gaussian distribution. The modeled
exceeds 100,000 draws. After the MCMCS is performed, the resulting normal
normal scoring.
28
Figure 3.3: Three-dimensional projections of a ten-dimensional Gaussian PDF
sampled with 1,000 (upper left), 10,000 (upper right), 100,000 (lower left) and
1,000,000 (lower right).
usually limited, rock physics modeling is used to populate the multivariate PDFs
with data or seismic lithofacies that are not sampled by the existing wells, but are
29
estimate the effective bulk modulus of a fluid filled porous rock when the original
( )
(3.1)
, (3.2)
where:
= Porosity,
shown in Figure 3.4. The shown PDF was generated from well log data and it
defines the relations among the rock properties for clean sandstone. This
30
Figure 3.4: 2D Projections of the multivariate probability distribution function. A)
Bivariate P-wave velocity versus S-wave velocity PDF; B) Bivariate P-wave
velocity versus bulk density PDF; C) Bivariate S-wave velocity versus bulk
density PDF; and D) Bivariate Porosity versus bulk density PDF.
properties when the pores are filled with gas. We estimated the acoustic
properties of the brine and gas using the Batzle and Wang empirical model
(Batzle and Wang, 1992) and assuming salinity of 15,000 ppm, temperature of
84°C, pore pressure of 25 MPa and gas specific gravity of 0.6. The irreducible
water saturation of the modeled sandstone was assumed 8%; hence the fluid
properties were calculated for a saturation of 92% gas and 8% brine using the
31
homogeneous saturation model (Mavko et al., 1998). The fluid properties at the
Table 3.2: Fluid properties generated using Batzle and Wang (1992) fluid
properties empirical relations.
expected, the replacement of the original fluid with less dense and more
compressible fluid will decrease the density and bulk modulus. In contrast, the S-
wave velocity slightly increases due to the density effect on the S-wave velocity
expression. The comparison of the elastic properties histograms for the original
and fluid replaced sands shows that the fluid substitution introduces more
uncertainty in the gas filled rock in comparison with the brine-filled rock (see
Figure 3.6). A similar observation was made done by Sengupta (2000) in her
32
Figure 3.5: 2D Projections of the modeled multivariate probability distribution
function. A) Bivariate P-wave velocity versus Vs PDF; B) Bivariate P-wave
velocity versus bulk density PDF; C) Bivariate S-wave velocity versus bulk
density PDF; and D) Bivariate Porosity versus bulk density PDF.
33
Figure 3.6: Elastic properties comparison between original saturation (left) and
gas replaced saturation (right).
This probabilistic rock physics modeling process can include any simple or
complex rock physics model as long as the rock property can be defined in the
multivariate PDF. For instance, the Dvorkin cemented sand model (Dvorkin and
Other applications of theoretical rock physics models are shown in Figure 3.7.
The effect of fluid replacement, adding cement or reducing porosity over the P-
wave velocity (Vp) are displayed in the porosity-Vp space. Figure 3.8 shows the
34
Figure 3.7: Example of theoretical rock physics models application to measured
P-wave velocity samples. The fluid replacement, cement modeling and porosity
perturbation effects are shown by red, green and black arrows.
35
3.7. Using MCMCS to Explore Model Space
function. The use of Monte Carlo simulation is central to the resolution of inverse
of model spaces on a regular grid is not feasible (since the required number of
When the MCMCS algorithm is well implemented, the shape of the target
PDF does not change with the number of simulated samples; this means that the
very similar to the original PDF, even though the number of simulated points is
more than one hundred times the number of original data points. We use the
Mean Squared Error (MSE) between the simulated and target distributions to
Figure 3.10 shows the MSE plot for the MCMCS example shown in Figure
3.9. For this particular MCMCS implementation the equilibrium is reached after
( ) ],
36
Figure 3.9: Comparison of an original bivariate distribution (left) and the MCMCS
distribution (right).
Figure 3.10: Mean-square-error plot for the MCMCS of the example shown in
Figure 3.9.
37
4. 1D MARKOV CHAIN STOCHASTIC SEISMIC INVERSION
4.1. Introduction
This chapter presents the 1D Markov Chain stochastic inversion algorithm. First,
it shows the general workflow for the 1D inversion scheme. Second, it describes
the parameters that describe a system (Tarantola, 2005). In contrast, the forward
problem involves the prediction of the measurements given the parameters that
describe the studied system (Figure 4.1). In Earth science the goal of the inverse
observations.
Figure 4.1: The inverse and forward problems are complementary mathematical
problems.
38
The content of this chapter was partially presented in the paper IPTC-18040 of the 2014 International Petroleum
Technology Conference in Kuala Lumpur.
This dissertation is focused on the specific problem of estimating the
earth’s parameters from surface seismic data. As with any inverse problem in
earth science, seismic inversion does not have a unique solution. Hence, we
generated from the available data and model. The probabilistic theory applied to
Chain stochastic simulation that uses statistical rock physics to constrain the
space of possible solutions. Mukerji et al. (2001) introduced the statistical rock-
introducing the inversion technique that directly produces rock property volumes
39
instead of elastic property volumes; using our methods, inversion results can be
inversion technique consists of four main steps: the rock properties scenario
generation, the synthetic data calculation and the comparison between the real
generated, and the model properties are calculated from the previously defined
produced using the elastic properties from the subsurface model. Finally, the
synthetic trace is compared to the real seismic trace; if the similarity between
both traces is larger than a given threshold, the model is accepted as a possible
solution and stored. In contrast, if the likeness between the real and the synthetic
traces is smaller than the mentioned threshold, the model is erased and a new
model is generated.
Therefore, the inversion outcome can be directly introduced into the geological
40
Figure 4.2: Simplified 1D stochastic inversion workflow.
layers geometry or how the seismic lithofacies are arranged within the studied
rock property distribution. The second is the rock properties of the analyzed
lithofacies. The scenario generator algorithm works on two levels. First the
routine generates the geometry of the probable earth models. Once the geometry
41
is built, we assign rock properties to each layer. The rock properties are assigned
illustrated in Figure 4.3. First, the fraction of the seismic lithofacies that constitute
the analyzed interval is generated. Once the 1D array with the fractions of the
seismic lithofacies that constitute the study interval are generated, the 1D array
and Frank Yates in 1938 (Fisher et al., 1938). The Fisher-Yates shuffle algorithm
unchosen numbers;
c) Counting from the low end, strike out the Xth number not yet selected and
store it;
d) Repeat from step 2 until all the numbers have been selected.
The 1D array of numbers stored after all the numbers are selected is now
42
An example of the application of the Fisher and Yates method, for an
Table 4.1: Example of the Fisher and Yates shuffle algorithm. The permuted
interval ranges from 1 to 5. The red cells correspond to the number that is
selected in each iteration.
4 1 1 2 3 4 5 41532
43
The generated seismic lithofacies indicator 1D array is the template used
for the rock property 1D arrays generation. The property 1D arrays are produced
in an order that starts with the lithology and porosity, and finishes with the elastic
properties. Each cell in the property 1D array is filled with the property
information calculated from the multivariate PDF and defined by the Markov
Chain. The accuracy of the rock properties model relies on the idea that the
a) Fluid contacts: For geological models with local fluid contacts, all samples
of gas reservoirs should be above the oil samples and all samples with oil
should be above the water samples.
posteriori PDF ( ( )) from the a priori PDF of the data ( ( )) and the model
( ( )). Equation 1.1 is valid when the information in the space of measured
parameters (data) has been obtained independently of the prior information in the
operator ( ( )) that predicts data (Equation 1.2) from the model. The predicted
44
values are not necessary identical to the observations since there are
convolutional model for the synthetic trace generation. This model is expressed
as follows:
( ) ( ) ( ), (4.1)
where:
The convolutional model of the seismic trace assumes that the latter is the
recorded wavelet (Sheriff, 1989). Since the elastic property 1D arrays are
generated in the depth domain, they have to be stretched to two-way time using
function is part of the input data. Another important part of the input data is the
the measured seismic data at the same angle ranges at which the synthetic data
45
After the elastic properties arrays are converted to two-way travel time,
they are interpolated to the same regular sampling as the input seismic data.
This is a very important process since the time sampling acts as a frequency
operator that defines the maximum vertical resolution of the inverted earth model
(see Figure 4.4). Finally, a Nyquist filter is applied to the time-stretched elastic
Figure 4.4: Comparison among the depth, time and regularized time 1-D P-wave
velocity arrays. The black circles show how the regularization in the time domain
can affect the elastic properties of the modeled subsurface.
Once the elastic properties arrays in time have been generated, the angle-
dependent reflectivity profiles are produced using the Aki and Richards linear
approximation (Aki and Richards, 1980) of the Zoeppritz equation. The Aki and
46
Richards linear approximation can be expressed in terms of P-wave velocity, S-
( ) ( ) [ ( )] ( ) (4.1)
where:
ΔVp=Vp2-Vp1 Vp=0.5(Vp1+Vp2)
ΔVs=Vs2-Vs1 Vs=0.5(Vs1+Vs2)
Δρ=ρ2-ρ1 ρ=0.5(ρ1+ρ2)
ϴ=0.5(ϴ1+ ϴ2);
Vp, Vs, ρ and ϴ are the P-wave velocity, S-wave velocity, bulk density and
incidence angle, respectively. The sub-indexes 1 and 2 refer to the upper and
( ) ( ) (4.2)
where A, B and C are the AVO intercept, AVO gradient and AVO curvature. We
used the first two terms of AVO Equation 4.2 for the estimation of the angle-
dependent reflectivity arrays. The two-term Aki and Richards AVO equation is
47
Figure 4.5: Schematic representation of the synthetic traces generation
algorithm.
the misfit function between the real and synthetic seismic traces. The synthetic
trace is compared to the real seismic trace and if the difference between them is
less than a given threshold the model is accepted; otherwise, the model is
cascade sequence that starts with the nearest and finished, with the farthest
angle-stacked traces. Once a trace is accepted the algorithm analyzes the next
angle trace until all the angle-dependent reflectivity traces are accepted. There
are many criteria to estimate the degree of similarity between the synthetic and
48
real traces; we used L2-norm to measure the similarity between the real and
√∑ ( ) , (4.2)
where: Ti Syn= Synthetic trace, Ti Real= Real trace, and i= ith Element of the
An example of one trace misfit analysis is shown in Figure 4.6. The top
graph shows the real (black) and synthetic (blue) traces. The difference between
the synthetic and real traces is displayed in the center graph. Finally, the
Figure 4.6: Misfit calculation example. In this synthetic example the cumulative
2
L -norm shows the intervals with larger misfit.
49
The probable models generated by the proposed inversion scheme are a
valid solution to the studied inverse problem; all of them honor the input rock
50
5. SYNTHETIC EXAMPLES
5.1. Introduction
of lithology and saturation from seismic and well log data. In this chapter we
present the results of the application of our proposed inversion scheme on two
different sets of synthetic data. The synthetic data sets were generated based on
real well data with the goal of testing the inversion in two different geological
environments. The first synthetic data set is composed of low porosity high
impedance sand. The goal of the second example is to replicate a high porosity
low impedance clastic environment. In this example, the target reservoir rock is
multivariate PDFs that were generated from the synthetic log and extended by
theoretical models. For instance, changes in pore fluid properties were modeled
using Gassmann fluid replacement theory. The rock’s properties in both synthetic
log data sets were based on real measured data. Next, the data space was
generated using synthetic seismic data modeled from synthetic elastic well logs.
As discussed in Section 4.4, the forward modeling operator was based on the
51
The content of this chapter was partially presented in the paper IPTC-18040 of the 2014 International Petroleum
Technology Conference in Kuala Lumpur.
5.2. Synthetic Test 1: Cemented Sand
The first synthetic data set was based on real well log data measured in a
Cretaceous interval; the in situ well logs are displayed in Figure 5.1. The
synthetic model was formed by two sands encapsulated into a shale sequence.
The top upper sand is located at 3,347 meters depth with a thickness of 40
meters. The lower sand is thinner with a thickness of 20 meters and located just
32 meter bellow the base of the upper sand. The target reservoirs are cemented
sand with average porosity of 16%. The background P-wave impedance is about
10 km*kg/sec*cc. The average P-impedance corroborates that the sands are well
Figure 5.1: Model 1 synthetic well logs showing (from left to right) facies
indicator, shale volume, P-wave velocity, bulk density, S-wave velocity, porosity,
P-wave impedance and S-wave impedance.
52
In this kind of elastic scenario, fluid variations have very small influence on
reservoir thickness and reservoir quality. In our study here, the gross reservoir
inverted interval. The reservoir quality is defined by two properties, which are the
fraction of shale (shale volume) and the total porosity of the sandstone. The
The elastic fluid properties profiles for brine, oil and gas were estimated
using the Batzle and Wang (1992) equations and assuming pressure and
temperature gradients of 9.72 MPa/km and 22.3 °C/km. Oil gravity, gas gravity
and salinity were 36° API, 0.6 and 8,000 ppm respectively. Figure 5.2 shows the
P-wave and density profiles for three possible pore fluids. Once the fluid
properties are estimated, the fluid replacement is performed using the Gassmann
relations. The fluid substitution was performed from brine to 90% hydrocarbon
saturated sands.
53
Figure 5.2: Synthetic model 1 pore fluids velocity and density profiles.
The final seismic lithofacies multivariate PDF projections for the Vp-Vs,
Vp-Rhob, and PhiT-Vp spaces are shown in Figure 5.3. The effect of the fluid
there was good lithology discrimination in all projected spaces. Hence, the test
was planned focus on the prediction of gross reservoir thickness and reservoir
quality.
54
Figure 5.3: Final seismic lithofacies multivariate PDF projections for the Vp-Vs
(Upper), Vp-Rhob (Center), and PhiT-Vp (Lower) spaces. Green, blue, brown
and red contours denote shales, brine sands, oil sands and gas sands,
correspondingly. Some empirical models are plotted for comparison including
Castagna’s mudrock line, Castagna’s 1993, Han’s 1986, Gardner’s sand,
Gardner’s shale, Voigt bound, Reuss bound and Nu’s modified Voigt bound for
quartz mixture.
logs converted to time. They showed strong positive reflection coefficients at the
top of both sands. It was clearly evident that one of the challenges in this test
was to accurately predict the reservoir geometry in the target interval. Figure 5.4
55
shows the model 1 angle dependent reflectivity profiles. It can be observed how
the reflectivity of the top and base of the sands decrease with the angle of
Hz (Figure 5.5).
56
Figure 5.5: Ricker wavelet with a central frequency of 25 Hz.
For this test we assumed the stationarity of the wavelet, meaning that
wavelet did not vary with incident angle or propagation time. The results of this
forward modeling are shown in Figure 5.6. A set of five angle traces were
generated equally spaced in an interval that ranged from 0° to 32°. This angle
gather was used as the measured seismic data for this synthetic test.
Figure 5.6: Synthetic angle gather generated from the synthetic model 1.
57
Once the “real traces” were generated the next step was to produce the
multivariate PDF for each of the analyzed lithofacies. The 2D projections of the
multivariate PDF (Figure 5.6) show that the sands have larger p-wave velocity,
lower Vp/Vs ratio and larger density than the surrounding shales. The
populations.
58
The inversion was performed for one million realizations using the model
shown in Figure 5.6 as the measured data and the multivariate PDFs shown in
figure 5.7 as the rock physics models. Modeled traces with similarity factor (α)
smaller than 0.9 were selected as possible solutions for the stochastic inversion
of the synthetic data. The comparison between the synthetic seismic data and
the 1,112 possible solutions revealed the close similarity between the synthetic
measured data and the probable models seismic traces. The residuals were
small, the time arrivals (Figures 5.8) and waveforms (Figures 5.9) similar.
Figure 5.8: Real seismic trace (left) and possible models traces (right)
2
comparison. The selected traces have L -norms smaller than 0.9.
59
Figure 5.9: Synthetic versus modeled waveform comparison for four possible
models with α<0.9.
The seismic lithofacies possible models with α<0.9 are displayed in Figure
5.10. It can be observed that most of the sand samples are located inside the
limits of the upper and lower sands. In fact, these sands are well-delimited by
Gross reservoir thickness is defined by the fraction of sand in the target interval.
Figure 5.11 shows the PDF of the sand fraction for all probable models with
60
α<0.9. The most expecting value of the sand fraction is 0.13, which is the same
fraction defined in the synthetic model. However, there is not a sole solution,
since the inversion produces a range of likely solutions that are constrained by
Figure 5.10: Probable models with α<0.9. The black, red, yellow and green lines
represent the top of upper sand, base of upper sand, top of lower sand and base
of lower sand, respectively.
61
Figure 5.11: Sandstone fraction for probable models with α< 0.9.
calculated from the probable models, we were able to define the key rock
properties for each lithofacies, and especially for the sandstones. For this
particular case, total porosity was a very important property for the definition of
reservoir quality. From the total-porosity models in Figure 5.12 we were able to
calculate some descriptive statistical coefficients; for this particular test, the most
likely sand total porosity was about 0.15, with a standard deviation of 0.7. From
the total porosity and volume of shale probable models (Figure 5.13), it was
possible to estimate the effective porosity of the target reservoir (Figure 5.14).
62
Figure 5.12: Total porosity of the probable models with α<0.9. The black, red,
yellow and green lines represent the top of upper sand, base of upper sand, top
of lower sand and base of lower sand, respectively.
Figure 5.13: Volume of shale probable models with α<0.9. The black, red, yellow
and green lines represent the top of upper sand, base of upper sand, top of lower
sand and base of lower sand, respectively.
63
Figure 5.14: Effective porosity probable models with α<0.9. The black, red,
yellow and green lines represent the top of upper sand, base of upper sand, top
of lower sand and base of lower sand, respectively.
Figure 5.15: Elastic models with α<0.9. The black, red, yellow and green lines
represent the top of upper sand, base of upper sand, top of lower sand and base
of lower sand respectively.
64
As has been mentioned before, all the solutions generated by the
seismic inverse problem. Therefore, the statistics of the solutions have more
significant relevance than the models themselves. The comparison between the
rock properties of the synthetic model and the solutions (Figure 5.16) shows a
very close similarity between the synthetic data and the inversion solutions. The
synthetic model used in this exercise is realistic and it is based on measured well
log information.
65
It is clear that if α decreases the similarity of the possible models will
increase. However, the number of possible models will decline as well. Figure
5.17 shows the possible solution models for α<0.045; in this particular case, we
obtained 78 possible models instead of the 1,112 calculated with α<0.9. This is
Figure 5.17: Probable models with α<0.045. The black, red, yellow and green
lines represent the top of upper sand, base of upper sand, top of lower sand and
base of lower sand, respectively.
The first synthetic test proved the viability of the application of the
proposed inversion scheme to predict the geometry and quality of the reservoir of
66
the target formation. The goal of the second synthetic test is to replicate a high-
31%. In this geological setting, fluid properties variations have very strong effect
amplitude reflections called “bright spots”. In this test, the objective was to define
the fraction of sand, the fraction of hydrocarbon sand, the fluid contact, the
The synthetic well logs, displayed in Figure 5.18, show the in situ case in
black line, brine case in blue and the gas case in red. The fluid replacement was
performed using the Gassmann relations. The elastic fluid properties profiles for
brine, oil and gas (Figure 5.19) were estimated using the Batzle and Wang
MPa/Km and 20 °C/Km. The gas gravity and salinity were 0.6 and 10,000 ppm
respectively. The well logs indicated that the brine-saturated sandstone had
larger impedance than the surrounding shale. However, when the pore fluid was
substituted from brine to a mixture of 90% gas and 10% brine, the sandstone’s
impedance became lesser than the shale’s acoustic impedance. Again, the rock-
physics diagnostic plot (Figure 5.20) shows the most important rock properties
for the three analyzed seismic lithofacies; in general, the elastic discrimination
67
1950
2000
2050
2100
2150
Depth [m]
2200
2250
2300
2350
2400
2450
0 0.5 1 0 0.2 0.4 1.5 2 2.5 1.5 2 2.5 0.5 1 1.5
Vshale [fraction] Porosity [fraction] Vp [km/sec] Rhob [g/cc] Vs [km/sec]
Figure 5.18: Model 2 synthetic well logs showing (from left to right) facie
indicator, shale volume, porosity, P-wave velocity, bulk density and S-wave
velocity. Black, blue and red lines correspond to in situ, brine and gas
saturations.
1950
2000
2050
2100
2150
DEPTH [m]
2200
2250
2300
2350
2400
2450
0.5 1 1.5 2 0 0.5 1 1.5
Vp [km/sec] Rhob [g/cc]
Figure 5.19: Synthetic model 2 pore fluids velocity (left) and density profiles
(right). Blue and red lines correspond to brine and gas, respectively.
68
Figure 5.20: Final seismic lithofacies multivariate PDF projections for the Vp-Vs
(Upper), Vp-Rhob (Center), and PhiT-Vp (Lower) spaces. Grey, blue, and red
contours denote shales, brine sands and gas sands, correspondingly. Some
empirical models are plotted for comparison, including Castagna’s Mudrock line,
Castagna’s 1993, Han’s 1986, Gardner’s sand, Gardner’s shale, Voigt bound,
Reuss bound and Nu’s modified Voigt bound for quartz mixture.
impedance contrasts in the studied interval; the top of the gas sand is delineated
positive reflection coefficient corresponds to the gas-brine fluid contact within the
sandstone. This seismic interface has a relative small positive AVO gradient.
69
Finally, the base of the brine-saturated sand is defined by a negative reflectivity
25 Hz Ricker wavelet (Figure 5.5) with the reflectivity logs displayed in Figure
5.22. As was expected, the most important reflection is the gas-brine fluid contact
functions were generated for the shale, the brine sand and the gas sandstone
70
Figure 5.22: Synthetic angle gather generated from synthetic model 2.
71
The inversion was five million realizations with the constraint that all brine
samples should be above the deeper gas sample. This is important because we
assumed a unique fluid contact in the whole interval; the models with L2-norms
smaller than 0.88 were selected as possible solutions for the inversion. The
comparison between the synthetic model and the 664 obtained solutions (Figure
5.24) shows a close similarity between the synthetic input data and the inversion
results. The waveform comparison for four inverted solutions and the synthetic
model (Figure 5.25) demonstrate that the main events are well reproduced by the
inversion algorithm.
0.08
2
0.06
2.05
2.1 0.04
Amplitude [Dimensionless]
2.15 0.02
Time [sec]
2.2 0
2.25 -0.02
2.3 -0.04
2.35 -0.06
2.4 -0.08
Figure 5.24: Real seismic trace (left) and possible models traces (right)
2
comparison. The selected traces have L -norms smaller than 0.88.
72
Figure 5.25: Synthetic versus modeled waveform comparison for four possible
models with α<0.88.
The seismic lithofacies distributions for the 664 solutions are shown in
large consistency for the depth of the top of the sand, fluid contact and base of
the sand. We generated histograms of the depth of the three seismic lithofacies
5.27).
Table 5.1 shows the comparison between the synthetic data and the
depths obtained by the inversion; these results demonstrate that the inversion
73
Figure 5.26: Probable models with α<0.88. The green, black and white represent
the shales, gas sands and brine sandstone seismic lithofacies, respectively.
350 350
300
300 300
250
250 250
200
Counts
Counts
Counts
200 200
150
150 150
100
100 100
50
50 50
0 0 0
2155 2160 2165 2170 2175 2190 2195 2200 2205 2210 2210 2220 2230 2240
Depth of Top of the Sand [m] Depth of Gas-Water Contact [m] Depth of Base of the Sand [m]
Figure 5.27: Depth of top of the sand (left), gas-water contact (center) and base
of the sand (right) histograms.
74
Table 5.1: Comparison between the synthetic data and the depths obtained by
the inversion.
Gas-Water
Top of Sand Base of Sand
Contact
Synthetic Model
2166.00 2202.00 2222.00
Value [m]
Mean [m] 2165.80 2198.70 2222.30
Median [m] 2166.00 2198.00 2222.00
Standard Deviation 2.82 2.63 3.21
Variance 7.97 6.92 10.34
Figures 5.28 to 5.31 show total porosity, shale volume, and effective
porosity, respectively. The effective porosity was calculated from the total
porosity and volume of shale properties. The elastic property solutions are shown
in figure 5.32. The fluid contact is easily observed in all the elastic models.
2000 0.32
2050 0.3
2150 0.26
Depth [m]
2200 0.24
2250 0.22
2300
0.2
2350
0.18
2400
0.16
100 200 300 400 500 600
Number of Traces
75
Volume of Shale Vertical Models
0.9
2000
0.8
2050
0.7
0.5
2200
0.4
2250
0.3
2300
0.2
2350
2400 0.1
0.3
2000
2050 0.25
2200 0.15
2250
0.1
2300
2350 0.05
2400
0
100 200 300 400 500 600
Number of Traces
Figure 5.30: Effective porosity probable models with α<0.88 calculated from the
total porosity and shale volume properties.
76
Figure 5.31: Elastic models with α<0.88.
5.4. Conclusions
inversion approach. The first example showed the capability of the novel
77
inversion scheme to estimate two critical rock properties for hydrocarbon
the gas-filled sandstone can be easily discriminated from the shales and brine-
saturated seismic lithofacies. We also showed the accuracy of the method in the
reflection seismic data, using Markov Chain Monte Carlos Simulation (MCMCS)
78
6. 1D REAL DATA APPLICATION
6.1. Introduction
inversion algorithm to a real data set. The data set includes a group of well logs
and a 2D line of prestack seismic data extracted from a 3D volume crossing the
marine sediments deposited on a platform margin. From well-log data and the
The potential reservoirs are unconsolidated sandstones with total porosity that
ranges from 35% to 39%.The pore space can be filled by gas, low saturation gas
seismic lithofacies: shales, gas-filled sands, low saturation gas sands (LSGS)
and brine sands. Figure 6.1 shows the in situ saturation logs for the two hundred
and twenty meter study interval. The target interval has two sands: the upper
sand is interceded with total thickness of twenty meters and a sand fraction of
73%; the lower reservoir is eighty meters of blocky sand with three fluid faces:
The seismic data included raw PSTM gathers with a time sampling of 2
msec. The near angle stack generated from 0° to 15° is shown in Figure 6.2; the
top of the strong seismic anomaly related to the top of the gas reservoir is
79
represented by large negative seismic amplitude. The top of the upper and lower
sands are located at 72 ms and 104 ms, respectively. The goal of this inversion
was to predict total porosity, volume of shale, effective porosity, and fluid type
Figure 6.2: Near angle stack; the top of the upper and lower sands are located at
72 ms and 104 ms, respectively.
80
6.2. Seismic Data Conditioning
noise and to improve the signal-to-noise ratio for the stochastic seismic inversion.
This was an important stage, since the forward modeling operator assumed that
all the reflections corresponded to primary reflections and random noise, and no
coherent noise was present in the seismic data. The seismic data was also
resampled to one millisecond to match the time sampling of the synthetic models.
The angle gather at Well-1 location is displayed in Figure 6.3; the amplitude
extraction at the top of the upper sand shows a strong AVO effect associated
Figure 6.3: Angle gather at the Well-1 location (Upper). The amplitude extraction
at the top of the upper sand shows a strong AVO effect (Lower).
81
After the seismic data was conditioned and converted to angle gathers,
the well logs were converted to time and the well-to-seismic tie was performed
(Figure 6.4). The well log tie shows close similarity between the zero offset trace
and the acoustic synthetic trace (right track). The angle-dependent extracted
wavelets are shown in Figure 6.5. The L2-norm calculated from the real and
with α≤0.88 was a possible solution to the studied real data application.
Figure 6.4: Well logs converted to time showing the well to seismic tie (right
track); the black line corresponds to the measured seismic trace and blue the
synthetic seismogram.
82
1
oo
0.8
8o
16o
0.6
24o
Amplitude [Dimensionless]
32o
0.4
0.2
-0.2
-0.4
-100 -80 -60 -40 -20 0 20 40 60 80 100
Time [ms]
PDF for the four observed seismic lithofacies: shale, gas sand, low gas
relation was performed to better understand the effect of fluid variations in the
target reservoir sands. Figure 6.6 shows the well logs for the in situ, 90% gas,
The 2D projections of the multivariate PDF (Figure 6.7) show the feasibility
properties. Only the low saturation gas sands cannot be easily discriminated from
gas sands.
83
Figure 6.6: Well 1 logs after fluid substitution. Black, red and blue lines
correspond to in situ, 90% gas and 100% brine-saturated sands.
Vs [km/sec]
Vs [km/sec]
Vs [km/sec]
1.2 1.2 1.2 1.2
1 1 1 1
2 2.5 3 2 2.5 3 2 2.5 3 2 2.5 3
Vp [km/sec] Vp [km/sec] Vp [km/sec] Vp [km/sec]
Rhob [g/cc]
Rhob [g/cc]
Rhob [g/cc]
2 2 2 2
Vp [km/sec]
Vp [km/sec]
Vp [km/sec]
2 2 2 2
0.2 0.3 0.4 0.2 0.3 0.4 0.2 0.3 0.4 0.2 0.3 0.4
PhiT [fraction] PhiT [fraction] PhiT [fraction] PhiT [fraction]
84
6.4. The Seismic Stochastic Inversion
The rock property models were generated assuming that all gas samples
were located above the low saturation gas sands and all LSGS were situated
above the brine sand. The fluid contacts can be situated at any depth as long as
the order of the fluid interfaces is followed. The properties of the seismic
lithofacies were assigned from the multivariate PDFs by using the MCMCS
engine to sample them. The proposed inversion algorithm generated data in two
domains; the first is the rock physics simulation domain that is generated in
depth. The second was the forward-modeling domain that was built in two-way
travel time. For this real data application, the vertical sampling of the rock physics
simulation grid (depth) was fixed to one meter and the two-ways travel time
sampling for the forward model grid was set to one millisecond. The depth and
time grids were selected to minimize the resampling effect on the well-log data.
Five million realizations were performed and 320 models with normalized
α≤0.88 were selected. The comparison between the real traces and the 320
solutions for 0° and 32° incident angles are shown in Figure 6.8.
85
Angle Traces at 0o
2
Amplitude [Dimensionless]
50 1
Time [msec]
100 0
150 -1
200 -2
50 100 150 200 250 300
Number of Traces
Angle Traces at 32o
2
Amplitude [Dimensionless]
50 1
Time [msec]
100 0
150 -1
200 -2
50 100 150 200 250 300
Number of Traces
Figure 6.8: Comparison between real angle traces (upper and lower left) and the
320 solutions for 0° (upper right) and 32° (lower right) incident angles.
Figure 6.9 shows that the similarity among the traces is large enough to
assume that all of 320 model traces were generated from probable geological
models. The upper and lower sands are well-delimited by most of the probable
inversion solutions. However, the gas-LSGS fluid contact can be located at any
low saturation gas sands. The possible models plot shows that all the possible
solutions for the LSGS-brine sand fluid contact have the fluid interface located in
the lower sand. In fact, most solutions have the fluid contact situated more than
86
Figure 6.9: Probable models with α≤0.88. Grey, black and white represent
shales, gas sands and low saturation gas sands, respectively.
effective porosity prediction from the seismic inversion, as has been mentioned.
Figures 6.10, 6.11 and 6.12 show the total porosity, volume of shale, and
effective porosity for the 320 solutions. The elastic models shown in Figure 6.13
demonstrate that there are almost no differences in the elastic properties of the
gas sands and low saturation gas sands. In fact, the main differences are found
87
Figure 6.10: Total porosity of the 320 probable models.
88
Figure 6.12: Effective porosity of the 320 probable models.
89
6.5. Analysis
The combination of the probable models for the target zone was used for
the analysis of rock properties. Figure 6.14 shows the probability of the sand plot;
larger values mean higher possibility of having a sand sample at a given depth. It
can be observed that the upper sand has less probability than the lower sand,
to mention that the larger the cutoff the smaller the sand fraction in the analyzed
interval; we chose a conservative value of 0.2 as the threshold for this real data
set. The comparison between the real and the estimated top values for
probability equal to or larger than 0.2 is shown in table 6.1. The comparison
confirms the accuracy of our proposed inversion scheme for lithology prediction.
90
20
40
60
80
100
Depth [m]
120
140
160
180
200
220
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Probability of Sand
Table 6.1: Comparison between the real and the estimated lithology top values
for probability larger than 0.2.
91
Based on the inversion solutions for total porosity and effective porosity,
we generated the most probable vertical plot for both rock properties (Figure
6.15). The comparison shows that the inverted porosities underestimated the
larger porosity values and overestimated the lower observed well log porosity.
This smoothing effect is stronger in the upper sand where extreme property
0 0
20 20
40 40
60 60
80 80
100 100
Depth [m]
Depth [m]
120 120
140 140
160 160
180 180
200 200
220 220
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
Total Porosity [fraction] Effective Porosity [fraction]
Figure 6.15: Total porosity and effective porosity comparison. The black lines
correspond to the well log information and the blue lines to the inverted rock
properties.
92
The most-likely fluid geometry (Figure 6.16) was generated from the 320
probable models section (Figure 6.9) and constrained by the most expected sand
profile (Figure 6.14). The comparison between the real fluid interfaces and the
inverted fluid contact depths (Table 6.2) shows a surprisingly close agreement
Figure 6.16: Most expected fluid seismic lithofacies (left) and probability of fluid
type (right).
Table 6.2: Comparison between the real and the estimated fluid contact depths.
93
Finally, the elastic properties probable models generated from the
stochastic inversion are compared with the Well-1 elastic logs in Figure 6.17. The
results of the inversion are remarkable similar to the well log measured data.
However, the S-wave velocity of the interbedded shales within the upper sand
section was not well predicted. These shales have a very low Vp/Vs ratio
0 0 0
20 20 20
40 40 40
60 60 60
80 80 80
Depth [m]
Depth [m]
120 120 120
Figure 6.17: Comparison between the most expected elastic models (blue line)
against well logs (black lines).
94
6.6. Conclusions
inversion are remarkable similar to the available well log measured data. It is
important to emphasize that Well-1 was not used to constrain the inversion; it
was only used to populate the model space. The mean of the inverted rock
properties tends to be smoother than the measured values; in fact, the most
‘smoothing’ effect is stronger in layers where extreme property values are further
The S-wave velocity of the shale member of the interbedded upper sand
lithofacies might be included in the model space. Finally, the proposed inversion
approach demonstrates the viability of the integration seismic data, well log
95
7. 3D REAL DATA APPLICATION
7.1. Introduction
using the proposed inversion algorithm. The real data seismic inversion is
performed using the same data set presented in the Chapter 6. The main
geostatistics that are used to create lateral continuity to the stochastic seismic
inversion outcomes. The second part of the chapter shows the application of the
Finally, we discuss the convergence criteria and the efficiency of the proposed
inversion algorithm.
96
linear unbiased estimators that predict the value of a property in any given
properties and their spatial distributions. All the Kriging algorithms assume that
the observed data are the result of a random process; then the objective of the
that the local mean may vary greatly over the studied area. In this case, the
∑ ( ) ( ) ( ) ( )
{ , (7.1)
∑ ( ) ( )
where:
h = uα-uβ,
distance to the location that is being estimated. The weights λα are calculated
from the covariance function that is generally derived from the semivariogram
97
model. The semivariogram model is a function that describes the degree of
difference between known values at two locations across the target area. The
process at two point pairs p1 and p2 only depends on the separation distance h
and not on the orientation of the two points. However, real data often show
There are two types of anisotropy. The simplest type occurs when the
covariance form is the same in all directions but the range S changes. In this
case, there is a single sill, but the semivariogram reaches the sill in a shorter lag
( ), (7.2)
where R is the ratio of between the larger r1and the lower seal r2:
. (7.3)
the Ordinary Kriging weights. The spherical semivariograms are defined as:
98
( ) ( ) ( ) , (7.4)
where h is the distance vector and a is a fitting constant defined from the
observed data.
constraint rock properties model. Figure 7.1 shows the simplified stochastic
between the Markov Chain simulation for rock properties or the geostatistical
estimation of rock properties. In this real test we inverted the 50 traces that
constitute the 2D seismic line (Figure 6.2) using the Well-1 as initial well data and
the statistical rock physics model shown. To estimate the covariance we used an
anisotropic spherical semivariogram with a 0.1 nugget, a sill of 0.9 and ranges of
(Figure 7.2). We chose the horizontal range as the maximum distance to accept
99
Figure 7.1: Simplified 3D stochastic inversion workflow.
Figure 7.2: Anisotropic semivariogram model with a 0.1 nugget, a sill of 0.9 and
ranges of 500 meters and 50 meters.
100
Figure 7.3 shows the evolution of the variance model for one of the
stochastic simulation outcomes. The order selection of the seismic traces plays
selection of the trace’s order we implemented the Fisher and Yates shuffle
Therefore, any trace selection path should be independent from the previous
realizations. The upper section of Figure 7.3 shows the initial variance model, it is
clear that the closer to the Well-1 the smaller the variance. The lower variance
values are situated at the well location where the variance becomes zero. The
second, third and fourth sections (from top to bottom) represent three more
iterations that produced three new wells. Every new iteration reduced the
variance model (reducing the uncertainty) until the variance model was zero at
The inversion was performed for the four seismic lithofacies defined in
Chapter 6: shales, gas sands, LSGS and brine sands. Figure 7.4 shows three
porosity, volume of shale, P-wave velocity, bulk density and S-wave velocity
inversion outcomes are shown in figures 7.5 to 7.9, respectively. All the shown
models are probable and similar. However, they have small variations in the
depth of the seismic lithofacies interfaces, especially, in the gas sands and LSGS
101
Figure 7.3: Variance model evolution for one of the stochastic simulation
outcomes.
Figure 7.4: Three possible stochastic models. Shales are represented by facies
1, gas sands by facies 2, LSGS by facies 3 and brine sands by facies 4.
102
Figure 7.5: Three possible stochastic total porosity models.
103
Figure 7.7: Three possible stochastic P-wave impedance models.
104
Figure 7.9: Three possible stochastic S-wave velocity models.
A set of four hundred possible outcomes was generated from the seismic
section. This set of seismic inversions was used to calculate the most expecting
values of the estimated rock properties for each sample of the 2D seismic
inverted vertical section. The variance sections were used to better understand
the uncertainty associated with the inversion outcomes. Figure 7.10 shows the
most expected section of total porosity (upper section) and its associated
variance (lower section). Well-1 measured total porosity log is located in the trace
number 37. It can be observed that there is good agreement between the
measured data and the inverted section. However, Well-1 has a larger vertical
105
resolution than the inverted traces. As we observed in Chapter 6, the median of
values and overestimate lower porosity values. The variance section shows that
the upper gas sand has larger estimation uncertainty. This is due to the
interbedded gas-shale geometry of the upper sand. The zero variance trace
corresponds to the Well-1 location. The same phenomenon is observed in all the
rock property vertical sections; the upper sand has larger estimation errors than
the lower sand (Figures 7.11 to 7.14). In the bulk density inverted section (Figure
7.13) the variance of the upper sand and the lower sand gas intervals has larger
estimation errors than those of the LSGS and brine sand. This is caused by the
Figure 7.10: Most expected total porosity model generated from 400 stochastic
models.
106
Figure 7.11: Most expected volume of shale model generated from 400
stochastic models.
Figure 7.12: Most expected volume of P-wave velocity generated from 400
stochastic models.
107
Figure 7.13: Most expected volume of density generated from 400 stochastic
models.
Figure 7.14: Most expected volume of S-wave velocity generated from 400
stochastic models.
108
7.4. Convergence Criteria
misfit threshold α. For this real data application we chose the L2-norm value
obtained from the synthetic tie as value of α. This selection was based on the
idea that the forward model operator is imperfect and cannot perfectly reproduce
real seismic data. Other sources of misfits, besides the forward modeling
the smaller α becomes, the more realizations are needed to obtain a statistically
randomly selected from the five million realizations performed for the real
example inversion. It can be observed that only a few hundreds models look like
Figure 7.15: Models randomly selected from the real dataset inversion.
109
The synthetic traces generated from the models shown in Figure 7.15 are
shown in Figure 7.16. For other shown models, the calculated seismic traces are
very noisy, without recognizable seismic interfaces. For the 150,000 traces
shown in Figure 7.16, only nine were selected as probable solutions. The
selection of α value smaller than the similarity obtained from the well log tie
would have deteriorated the inversion outcome, because the possible solutions
Figure 7.16: Comparison between real angle trace (right) and 150,000 solutions
(left) randomly selected from the five million realizations performed for the real
example inversion.
110
7.5. Algorithm Efficiency
personal computer with two processors and eight cores. The results showed that
the amount of resources (computing time) needed by the our proposed algorithm
increase linearly (O(n)) with the number of tested models. The average running
Figure 7.17: Algorithm efficiency, the proposed algorithm increase linearly with
the number of tested models.
111
6.7. Conclusions
case studies was demonstrated. The outcome of the inversion was a set of key
rock property sections that can be directly used for the evaluation of the
inversion solutions were generated directly in the depth domain. Therefore, with
lateral continuity. As we observed in the 1D real data application, the real data
well log data. However, the median of the inverted traces has lower vertical
variability than the Well-1 logs. Larger estimation errors were found in the upper
The real-data tests showed that the similarity factor α is the most sensible
than the similarity obtained from the well log tie will deteriorate the inversion
outcome, since the possible solutions will be generated from a partially sampled
model space.
112
REFERENCES
Aki, K., and Richards, P. G., 1980, Quantitative seismology: theory and methods:
W. N. Treeman & Co. pp. 119-149.
Avseth, P., 2000, Combining rock physics and sedimentology for seismic
reservoir characterization in North Sea turbidite systems: Ph.D. dissertation,
Stanford Univ. pp. 14-49.
Avseth, P., Mukerji, T., and Mavko, G., 2005, Quantitative seismic interpretation:
applying rock physics tools to reduce interpretation risk: Cambridge University
Press. pp. 42-47.
Banchs, R. and Michelena, R., 2000, Well log estimates and confidence intervals
by using artificial neural networks: 70th Ann. Internat. Mtg,. Soc. of Expl.
Geophys., Expanded Abstracts, 1430-1432.
Batzle, M., and Wang., Z., 1992, Seismic properties of pore fluids: Geophysics,
57, 1396-1408.
Berryman, J. G., Berge, P. A., Bonner, B. P., 2002, Estimating rock porosity and
fluid saturation using only seismic velocities: Geophysics, 67, 391-404.
Bortoli, L. J., Alabert, F. A., Haas, A., and Journel, A. G., 1993, Constraining
Stochastic Images to Seismic Data, in A. Soares ed., Geostatistics Tróia 1992,
proc. 4th Inter. Geostat. Congr., Kluwer Academic Publ., 325-337.
Castagna, J. P., Batzle, M. L., and Eastwood, R. L., 1985, Relationships between
compressional-wave and shear-wave velocities in clastic silicate rocks:
Geophysics, 50, 571-581.
113
Castagna, J. P., Batzle, M. L., and Kan, T. K., 1993, Rock physics – The link
between rock properties and AVO response, in J. P. Castagna and M. Backus,
eds., Offset-dependent reflectivity – Theory and practice of AVO analysis: Soc. of
Expl. Geophys., 135-171.
Connolly, P., 1998, Calibration and inversion of non-zero offset seismic: 68th
Ann. Internat. Mtg,. Soc. of Expl. Geophys., Expanded Abstracts, 182-184.
Connolly, P., 1999, Elastic impedance: The Leading Edge, 18, 438-452.
Deutsch C., and Journel, A., 1998, GSLIB: Geostatistical Software Library:
Oxford University Press, 2nd edition.
Dvorkin, J., and Nur, A., 1996, Elasticity of High-Porosity Sandstones: Theory for
Two North Sea Datasets, Geophysics, 61, 1363-1370.
Dvorkin, J., Nur, A., and Yin, H., 1994, Effective Properties of Cemented
Granular Materials, Mechanics of Materials, 18, 351-366.
Fisher, A., Yates, F., 1948, Statistical tables for biological, agricultural and
medical research: Oliver & Boyd. Pp. 26-27.
Florez, J., 2005, Integrating geology, rock physics, and seismology for reservoir
quality prediction: Ph.D. dissertation, Stanford Univ. pp. 13-44.
Gardner, G., Gardner, L., and Gregory, A., 1974, Formation velocity and density
– The diagnostic basic for stratigraphic traps: Geophysics, 39, 770-780.
Gassmann, F., 1951, Über die elastizität poröser medien: Veirteljahrsschrift der
Naturforschenden Gesellschaft in Zürich, 96, 1-23.
Gonzalez, E. F., Mukerji, T. and Mavko, G., 2000, Facies identification using P-
to-P and P-to-S AVO attributes: 70th Ann. Internat. Mtg,. Soc. of Expl. Geophys.,
Expanded Abstracts, 98-101.
Goodway, B., Chen, T. and Downton, J., 1997, Improved AVO fluid detection and
lithology discrimination using Lame petrophysical parameters; “λρ”, “μρ”, & “λ/μ
114
fluid stack”, from P and S inversions: 67th Ann. Internat. Mtg,. Soc. of Expl.
Geophys., Expanded Abstracts, 183-186.
Grechka, V., 2008, Fluid substitution in porous and fractured solids: The non-
interaction approximation and Gassmann theory: International Journal of
Fracture, 148, 103-107.
Gray, D., 2002, Elastic inversion for Lamé parameters: 72th Ann. Internat. Mtg,.
Soc. of Expl. Geophys., Expanded Abstracts, 213-216.
Gutierrez, M., 2001, Rock physics and 3-D seismic characterization of reservoir
heterogeneities to improve recovery efficiency: Ph.D. dissertation, Stanford Univ.
Han, D., 1986, Effects of porosity and clay content on acoustic properties of
sandstones and unconsolidated sediments: Ph.D. thesis, Stanford University.
Harris, D. A., Lewis, J. J. M., and Wallace, D. J., 1993, The identification of
lithofacies types in geological imagery using neural networks, Conf. papers vol.,
EUROCAIPEP 93, Aberdeen.
Hashin, Z., and Shtrikman, S., 1963, A variational approach to the elastic
behavior of multiphase materials: J. Mech. Phys. Solids, 11, 127-140.
Hastings, W.K., 1970, Monte Carlo sampling methods using Markov Chains and
their applications: Biometrika, 57, 97–109.
Hearst, J. R., Nelson, P. H., Pailet, F. L., 2000, Well logging for physical
properties: McGraw-Hill, 2nd edition.
115
Hirsche, K., Boerner, S., Kalkomey, C. and Gastaldi, C., 1998, Avoiding pitfalls in
geostatistical reservoir characterization: A survival guide: The Leading Edge, 17,
493-504.
Landrø, M., 1999, Discrimination between pressure and fluid saturation changes
from time lapse seismic data, 69th Ann. Int. Meet., Soc. Expl. Geophys.,
Expanded Abstracts, 1651-1654.
Mallick, S., 2001, Hybrid inversion, elastic impedance inversion, and prestack
waveform inversion; 71st Ann. Internat. Mtg: Soc. of Expl. Geophys., 706-709.
Marion, D., 1990, Acoustical, mechanical and transport properties of sediments
and granular materials: Ph.D. thesis, Stanford University.
Mavko, G., and Jizba, D., 1991, Estimation grain-scale fluid effects on velocity
dispersion in rocks: Geophysics, 56, 1940-1949.
Mavko, G., Mukerji, T., and Dvorkin, J., 1998, The rock physics handbook:
Cambridge University Press.
Metropolis, N., Rosenbluth, A.W., Rosenbluth, M.N., Teller, A.H., Teller, E., 1953,
Equations of state calculations by fast computing machines: Journal of Chemical
Physics, 21, 6, 1087–1092.
Mukerji, T., Jørstad, A., Mavko, G., and Granli, J., 1998, Near and far offset
impedances: Seismic attributes for identifying lithofacies and pore fluids:
Geophysical Research Letter, 25, 4557-4560.
Mukerji, T., Avseth, P., Mavko, G, Takahashi, I., and Gonzalez, E. F., 2001,
Statistical rock physics: Combining rock physics, information theory, and
geostatistics to reduce uncertainty in seismic reservoir characterization: The
Leading Edge, 20, 313-319.
116
Oldenburg, D. W., Scheuer, T., and Levy, S., 1983, Recovery of the acoustic
impedance from reflection seismograms: Geophysics, 48, 1318-1337.
Ostrander, W. J., 1984, Plane-wave reflection coefficients for gas sands at non-
normal angles of incidence: Geophysics, 49, 1637-1648.
Possato, S., Saito, M., Curtis, M. P., and Martinez, R. D., 1983, Interpretation of
three-dimensional seismic attributes contributes to stratigraphic analysis of
Pampo oil field: 53rd Ann. Internat. Mtg,. Soc. of Expl. Geophys., Expanded
Abstracts, session: S16.2.
Ruiz, F., 2009, Porous grain model and equivalent elastic medium approach for
predicting effective elastic properties of sedimentary rocks: Ph.D. dissertation,
Stanford Univ.
Sengupta, M., 2000, Integrating rock physics and flow simulation to reduce
uncertainties in seismic reservoir monitoring: Ph.D. dissertation, Stanford Univ.
Taner, M. T., Koehler, F., and Sheriff, R. E., 1979, Complex seismic trace
analysis: Geophysics, 44, 1041-1063.
Tarantola, A., 2005, Inverse problem theory and methods for model parameter
estimation: Society for Industrial and Applied Mathematics.
117
Tonn, R., 1992, Statistical approach to correlate reservoir parameters and 3-D
seismic attributes, 62nd Ann. Internat. Mtg,. Soc. of Expl. Geophys., Expanded
Abstracts, 272-274.
Wu, Y., 2000, Estimation of gas saturation using P-to-S converted waves; 70th
Ann. Internat. Mtg,. Soc. of Expl. Geophys., Expanded Abstracts, 158-161.
Wyllie, M. R., Gregory, A. R., and Gardner, L. W., 1956: Elastic wave velocities in
heterogeneous and porous media, Geophysics, 21, 41-70.
Yin, H., Mavko, G., and Nur, A., 1993, Critical porosity; a physical boundary in
poroelasticity: 34th U. S. symposium on Rock mechanics, in Haimson, B. C., ed.,
Rock mechanics in the 1990s, Oxford-New York, International Journal of Rock
Mechanics and Mining Sciences & Geomechanics, Abstracts, 30, 805-808.
Zeng, H., Backus, M. M., Barrow, K. T. and Tyler, N., 1996, Facies mapping from
three-dimensional seismic data: Potential and guidelines from a Tertiary
sandstone-shale sequence model, Powderhorn Field, Calhoun County, Texas:
AAPG Bull., 80, 16-46.
118